Implement std::hash::Hasher trait on Digest

[rpmoore]

This implements the std::hash::Hasher trait on Digest which allows users of the library to use crc64fast in places where std::hash::Hasher is used. While users of this can use the Newtype pattern and implement this themselves, by implementing this trait it makes it more accessible to users of the crate.

[kennytm]

Hi. Before we merge or close this can you provide where is that std::hash::Hasher used? AFAIK cryptographic hash algorithms don't implement std::hash::Hasher but rather digest::Digest.

The main purpose of choosing to use CRC is to be compatible with existing protocols / file format that already uses this for calculating checksum.

OTOH, the Hasher trait is mainly used for HashMap/HashSet. The main merit of a hash function used in hash table is how well it can quickly distinguish different keys within the table's lifetime, rather than producing a stable result for a given input.

I don't think CRC makes a good hash table algorithm

• in terms of security, it is very simple to produce colliding keys
• in terms of speed, I don't think it can compete with FNV (fxhash) or ahash etc
• additionally, keys used for hash tables are often pretty short (so that hash calculation can still be treated as an O(1) operation), but crc64fast is designed for processing long input strings.

(Additionally our reference implementation crc v3 did not implement Hasher either. crc32fast did implement Hasher since the beginning but a rationale was not given.)

[rpmoore]

I'm using the trait to checksum files (or portions of files) to detect bit rot or mutations without using a cryptographically secure hash for speed purposes. I was following what crc32fast had done as it made being able to choose between either of them easier without having to wrap both in my own type. Reading the Hasher trait docs seem to imply it being for generic hashing of byte streams, so I didn't think it improper to implement, but if it's intended usage is only for Hash Tables then I agree CRC is not a good choice for Hash Tables. If not the Hasher trait though, is there a better trait to represent checksumming? A quick search doesn't surface anything that's an obvious choice.